BIT @ TRECVid SED 2013 Yicheng Zhao, Binjun Gan, Shuo Tang, Jing Liu, Xiaoyu Li, Yulong Li, Qianqian Qu, Xuemeng Yang, Longfei Zhang Key Laboratory of Digital Performance and Simulation Technology, Beijing Institute of Technology
Acknowledgement • Support by – Lab of Digital Performance and Simulation Technology • Reference – System Framework: [Informedia@tv11] – MoSIFT feature: [Chen09] – STIP feature: [Laptev05]
Background • First participation to TRECVid • Limited submission results – ObjectPut • No interaction • Focus on Location Information in feature-level
Outline • Framework • Motivation • Feature fusion • Parameter tuning • Experiments • Conclusion
Framework • Informedia@tv11
Framework • No Hot region detection • Only SVM with X^2 kernel X X SVM Chi-Square kernel
Framework Feature fusion • No Hot region detection with absolute location • Only SVM with X^2 kernel X X SVM Chi-Square kernel
Outline • Framework • Motivation • Feature fusion • Parameter tuning • Experiments • Conclusion
Motivation • Location invariance property of feature, e.g. MoSIFT, STIP, etc. – While TRECVid events are location related. • Normal Solution: Spatial Bag-of-Word • Why not add location information to the features?
About location information • Two kinds – Global absolute location (location of event) – Object based relative location • The location of the movement of the object part • Scale-invariant
Why absolute location ? • Relative location calculation depends on segmentation algorithm – Existing algorithm are not acceptable • Absolute location can transformed to relative location • No published conclusion – about feature- level absolute location’s Performance for Action Detection in Surveillance video
Outline • Framework • Motivation • Feature fusion • Parameter tuning • Experiments • Conclusion
Feature fusion • Spatio-temporal Feature (MoSIFT/STIP) • Absolute location of Feature (X,Y)
Feature fusion • Spatio-temporal Feature (MoSIFT/STIP) • Absolute location of Feature (X,Y) 256 Dim MoSIFT descriptor
Feature fusion • Spatio-temporal Feature (MoSIFT/STIP) • Absolute location of Feature (X,Y) ( X, Y ) x y , 0,1
Feature fusion • Spatio-temporal Feature (MoSIFT/STIP) • Absolute location of Feature (X,Y) 256 Dim MoSIFT descriptor + ( X, Y ) * x y , 0,1 Extend Spatio-temporal feature descriptor + ( X, Y ) * x y , 0,1
Outline • Framework • Motivation • Feature fusion • Parameter tuning • Experiments • Conclusion
Parameter tuning • Evaluate the Influence of beta in Action Recognition Spatio-temporal feature descriptor + ( X, Y ) *
Parameter tuning – Exp. Setting • PUMP dataset • 4 Fixed Cameras in different direction • “above”: 84 sequences, 6 people, 6 events 1 poweron/poweroff 2 caparm/cappump/ openpump/openarm 3 connect/disconnect 4 cleanpump/cleanarm 5 pushbutton 6 flushgreen/flushyellow Visualization of the MoSIFT feature point of 6 events *http://lastlaugh.inf.cs.cmu.edu/MedDeviceAssistance/downloads.html
Parameter tuning – Exp. Setting • Turning: x , 0,7 x • Measure: Cross validation, F1-Score • Spatial Constrain MoSIFT (SC-MoSIFT) + BoF
Parameter tuning – Beta
Parameter tuning – Best Beta Best value of Beta MoSIFT: 10^3
Parameter tuning – Best Beta Best value of Beta MoSIFT: 10^3 STIP: 10^0.7
Parameter tuning – Best Beta • Best Beta is influenced by the Avg. distance between two points of Spatio-temporal feature MoSIFT STIP Avg. distance between 10^3 10^1 two points
Parameter tuning – Best Beta • Beta is determined by the Avg. distance between two Spatio-temporal feature MoSIFT STIP Avg. distance between 10^3 10^1 two points Best value of Beta MoSIFT: 10^3 STIP: 10^0.7
Parameter tuning – Analysis • new features (SC feature) will be processed by K-means Feature fusion Visual vocabulary *The same setting with K-means informedia@tv11 (k=3000)*
Parameter tuning – Analysis • Beta influence the distribution of feature for clustering • Adding location information to visual vocabulary Concentrate together Spread out in space Distribution of clusters’ centers,(a)beta = 1, (b)beta = 1000
Results on PUMP • Better results on PUMP dataset – 15% improvement in F1-Score Result on PUMP “above” dataset Feature F1-Score SC-MoSIFT 0.7858 MoSIFT 0.6784
Results on PUMP • Evaluated the effectiveness of Spatial BoF Result on PUMP “above” dataset Feature F1-Score MoSIFT + Spatial BoF 0.74 SC-MoSIFT + BoF 0.78
Results on PUMP – Analysis • Two inspirations – Location Information in low-level-feature is efficient on classifying location related events – The location information in low-level-feature can achieve a better performance than in high-level-feature • Limitation of PUMP dataset – Main body in camera is static – relative location and absolute location are almost the same • Need more experiments
Outline • Framework • Motivation • Feature fusion • Parameter tuning • Experiments • Conclusion
Experiment on TRECVid • Similarity between PUMP and SED – Fixed camera – Event related to location = ObjectPut in CAM3
Experiment 1 – Setting • Submitted (BIT_2) • Event: ObjectPut • Training set: dev08 + eval08 • Setting: Comparing with Informedia@tv11 BIT_2 Informedia@tv11 SC-MoSIFT MoSIFT visual vocabulary size = 3000 visual vocabulary size = 3000 Spatial BoF with different frame Spatial BoF division method - Hot Region Detection SVM with Chi-Square kernel Cascade SVM
Experiment 1 – Results • Comparison with the Informedia@tv11 in MinDCR ObjectPut 2011 infomedia 1.0003 2013 BIT_2 1.0000
Experiment 1 – Analysis • Weaker classifier and no Hot Region Detection • But comparable result in MiniDCR – SC-MoSIFT may works • More control experiments are needed
Experiment 2 – Setting • Post-submission • Event: PersonRun • Training set: CAM3 in (dev08 + eval08) • Measure: cross validation, f1-score Run_1 Run_2 SC-MoSIFT MoSIFT visual vocabulary size = 3000 visual vocabulary size = 3000 Spatial BoF Spatial BoF SVM with Chi-Square kernel SVM with Chi-Square kernel
Experiment 2 – Results • F1-Score of PersonRun on CAM3 Feature F1-Score SC-MoSIFT 0.134783 MoSIFT 0.183908
Experiment 2 – Analysis • SC- MoSIFT’s performance depends on events – it not work on the detection of PersonRun
Experiment 2 – Analysis • Difference between PersonRun and ObjectPut – ObjectPut occurs in some particular locations – PersonRun occurs in a wide locations • The wide location result in bad visual vocabulary • The adaptive parameter is necessary
Outline • Framework • Motivation • Feature fusion • Parameter tuning • Experiments • Conclusion
Conclusion • This years TRECVid results show the great potential of feature fusion with location information.
Future work • Participate in next year’s SED, and test on more events with different fusion methods.
Thank you
Recommend
More recommend